AI Paper Atlas

Research map / Optimization / Distributed and compressed

Fine-tuning Language Models over Slow Networks using Activation Quantization with Guarantees

Jue WANG, Binhang Yuan, Luka Rimanic, Yongjun He, Tri Dao, Beidi Chen, Christopher Re, Ce Zhang

NeurIPS 2022 · Optimization › Distributed and compressed · approach: Gradient quantization and sparsification · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
compress activation changes (AQ-SGD) instead of activations directly (DirectQ) (ablation)activation quantization · QNLI/CoLA/WikiText2/arXiv, fw2-fw4convergenceImproved
use AQ-SGD instead of FP32 in slow networkspipeline-parallel fine-tuning · GPT2-1.5B, slow networksend-to-end speedup: 4.3×Improved
AQ-SGD versus DirectQ throughputactivation quantization · GPT2-1.5B, 100 Mbpstraining throughput: 3.5 -> 3.4No clear effect
AQ-SGD fw4 bw8 versus DirectQ fw4 bw8 throughputactivation quantization · GPT2-1.5B, 10 Gbpstraining throughput: 4.1 -> 4.0No clear effect
quantize activations with AQ-SGD instead of FP32pipeline-parallel training · GPT2-1.5B, 100 Mbpstraining throughput: 0.5 -> 3.4Improved
quantize activations with AQ-SGD instead of FP32pipeline-parallel training · GPT2-1.5B, 10 Gbpstraining throughput: 3.8 -> 4.0No clear effect
reduce network bandwidth 100x (10Gbps to 100Mbps)AQ-SGD · GPT2-1.5Btraining speed: 1.18× slowerHurt
combine AQ-SGD with QuantizedAdam gradient compressionend-to-end communication compression · GPT2-1.5B with data parallelismend-to-end speedup: 4.9×Improved
combine activation and gradient compression versus no compressionAQ-SGD + QuantizedAdam · GPT2-1.5B throughputthroughput: 8.5×Improved
disable either activation or gradient compression (ablation)AQ-SGD + QuantizedAdam · GPT2-1.5B throughputthroughputHurt
combine DirectQ with gradient compressionDirectQ · GPT2-1.5BconvergenceHurt
lower bandwidth increases communication timeAQ-SGD fw4 bw8 · GPT2-1.5B backward commcomm time per micro batch: 25 ms -> 125 msHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy